Google finished the global rollout of two Search Console features on August 31, 2026: the generative AI performance report, and the generative AI search control. The report tells you how often your links show up inside AI answers. The control lets you stop them showing up at all. One of those is useful to every site on the web. The other is a loaded gun with a very short list of legitimate targets.
This did not arrive because Google wanted to hand publishers a kill switch. The Competition and Markets Authority in the UK made it mandatory in June 2026, implementation began June 17, wider availability followed in July, and the worldwide switch happened at the end of August. Knowing that the control is a regulatory artifact rather than a product decision tells you something useful about how it will evolve: the roadmap dates are compliance deadlines, not feature ambitions.
What the generative AI search control actually does
It sits in Search Console under Settings, on a page called Search generative AI. Flip it, and two things happen to every URL in the property. Your links stop appearing in the covered surfaces, and your content stops being used as an input to generate the answers on those surfaces. That second half is the part teams miss. You are not just declining the citation. You are declining to be part of the source set that shapes what the engine says about your category.
Google puts the propagation window at one to two days for most content, with a few days for the tail, and the setting is reversible whenever you like. Child properties can inherit the parent setting or override it, which is the only granularity available right now and the reason large sites with a single property are stuck with an all or nothing call.
| QUESTION | ANSWER | WHY IT MATTERS |
|---|---|---|
| Does it change standard Search rankings | No. Google states it is not used as a ranking or inclusion signal outside generative AI features | Removes the main objection people raise, and shifts the decision onto pure economics |
| Which surfaces are covered | AI Overviews, AI Mode, and generative AI features in Discover | Third-party engines like ChatGPT and Perplexity are untouched by this setting |
| Is it page-level | No. Property-level only, with inheritance for child properties | A single property covering docs, blog and product means one decision for all three |
| Does it stop AI model training | No. That is Google-Extended, a separate control | Teams flip this expecting a training opt-out and get neither the outcome nor the visibility |
| How fast does it apply | Most content within one to two days, some longer | Fast enough to test, slow enough that a panic reversal is not instant |
| Is it reversible | Yes, at any time | The cost of a wrong call is measured in weeks of lost impressions, not permanence |
| Does opting out preserve traffic | No. Google states opted-out sites receive no traffic or impressions from generative features | There is no partial credit version of this decision |
The four things the toggle does not do
Most of the bad decisions we expect to see this quarter come from a mismatch between what the switch is for and what someone in a leadership meeting believes it is for. Four misreadings account for nearly all of them.
That fourth point deserves the most weight, because it is the one dressed up as strategy. Collective action by a publishing bloc is leverage. One enterprise site quietly flipping a Search Console setting is a rounding error, and the only party who reliably notices is the team that later has to explain a decline in top-of-funnel impressions.
The decision table
There are real cases for opting out. They are narrower than the discourse implies, and they share a shape: the content is the product, the answer substitutes for the visit, and the site has an existing revenue relationship that does not depend on discovery.
Illustrative decision weighting for four site profiles, scored on how much of the value is captured before the click. Higher means more of the value is consumed inside the AI answer itself. Illustrative scenario, not measured data.
A reference site whose entire value is a fact someone wanted has already been paid out when the answer renders. A B2B software company whose buyer takes four months and six stakeholders to make a decision has not. The gap between 88 and 22 in that scenario is the whole argument, and it is why the publisher debate and the B2B debate keep talking past each other.
Why the reporting cannot yet justify an opt-out
The generative AI performance report gives you impressions. It does not give you clicks, and it does not give you position. Under the UK timeline, click data is required by December 2026, which means anyone deciding today is deciding without the single number that would settle it: what an appearance in an AI answer is actually worth in sessions.
There are three more counting quirks worth knowing before you build a case on this data. Multiple results from the same site inside one generative feature count as a single impression, so a page that gets cited three times in one answer looks identical to a page cited once. The report inherits the standard 1,000 row limit, which truncates long tail analysis on any site of size. And the newest days arrive preliminary and marked with dotted lines, so a Monday morning read of the weekend is always an undercount.
| METRIC YOU WANT | AVAILABLE TODAY | WORKAROUND UNTIL DECEMBER |
|---|---|---|
| Impressions in AI Overviews and AI Mode | Yes, by page, country, device and date | None needed, this is the reliable part |
| Clicks from generative AI features | No | Segment landing page sessions against the impression trend and treat the correlation as directional |
| Position or citation rank inside the answer | No | Run a prompt panel on your top commercial questions and log which sources get named |
| Conversion value of an AI-sourced session | No | Tag and model it downstream, the same way you already handle assisted conversions |
| Which query produced the impression | Not in the generative report | Cross-reference the standard performance report on the same pages and dates |
None of that makes the report useless. It makes it a visibility instrument rather than a revenue instrument, which is exactly how we treat citation data generally. The measurement pattern here is the same one we described for tracking citation shifts across model releases: you are watching a trend line for shape, not reading a P and L.
“Withdrawing from a surface because you cannot measure it is not caution. It is choosing a smaller business with a tidier dashboard.”
There is also a real placement question underneath this. NewzDash found that nearly one in six US trending news queries, 15.5 percent, now place Top Stories inside AI Overviews, with the UK figure at 17.46 percent. For a news operation, opting out does not just remove an AI citation. It can remove a carousel placement that used to sit in a straightforward organic slot. Rahul Jain of Noble put the publisher case bluntly: opting out of AI Overviews will hurt most publishers more than it helps. Search Engine Journal collected those figures and reactions as the control reached general availability.
What to do instead of flipping the switch
For roughly 95 percent of the sites reading this, the correct action this month is to open the report, not the control. Establish the baseline while the data is young, because the first thing you will want in March 2027, when page-level controls arrive and the decision gets genuinely interesting, is six months of history showing which sections earn AI impressions and which do not.
Then work the supply side. A page that is easy to quote correctly earns citations across every engine at once, and the structural work that wins an AI Overview is the same work that wins a Perplexity citation. That starts with machine-readable structure, which we set out in the enterprise standard for structured data in AI search, and it compounds in a way that a defensive setting never will. Our generative engine optimization work is built almost entirely on that premise, and the pattern held across the Zenity engagement, where the same structural changes moved both classic rankings and citations in four engines.
One more piece of housekeeping while you are in Settings. Check what Google-Extended is set to, because that is the control most teams believe they have already configured and most have not. The two settings answer different questions and a company can reasonably want opposite answers to them: appear in the search-time answers, decline to feed the training corpus. You cannot get that combination by guessing which toggle is which.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.